A Conversational Question Answering on Phrase Ellipsis Recovery

Baseline

If you want to train your own data, please follow the glove and spacy, and generate your own features.

Base projects	flowQA	coQA	DrQA	QuAC (dataset)

Modified projects	flowQA_4_PERQA	coQA_4_PERQA	QuAC_4_PERQA	DrQA_4_PERQA

QuAC

parameters:

gpu5 quac_perq0:
    CoVe_opt=1, 
    MTLSTM_path='glove/MT-LSTM.pth', 
    answer_type_num=1, 
    batch_size=3, 
    concat_rnn=False, 
    cuda=True, 
    deep_att_hidden_size_per_abstr=250, 
    deep_inter_att_do_similar=0, 
    dev_dir='QuAC_data/', 
    do_hierarchical_query=True, 
    do_not_save=False, 
    do_prealign=True, 
    do_ptr_update=True, 
    do_residual=False, 
    do_residual_everything=False, 
    do_residual_rnn=False, 
    do_seq_dropout=True, 
    dropout_emb=0.4, 
    elmo_batch_size=12, 
    elmo_lambda=0.0, 
    epoches=30, 
    eval_per_epoch=1, 
    explicit_dialog_ctx=2, 
    final_output_att_hidden=250, 
    fix_embeddings=False, 
    grad_clipping=10, 
    hidden_size=125, 
    learning_rate=0.1, 
    log_file='output.log', 
    log_per_updates=20, 
    max_len=35, 
    model_dir='models', 
    momentum=0, 
    my_dropout_p=0.4, 
    name='', 
    ner_dim=8, 
    ner_size=19, 
    no_dialog_flow=False, 
    no_em=False, 
    optimizer='adamax', 
    pos_dim=12, 
    pos_size=51, 
    prealign_hidden=300, 
    prealign_option=2, 
    pretrain='', 
    ptr_net_attn_type='Bilinear', 
    ptr_net_indep_attn=True, 
    question_merge='linear_self_attn', 
    question_normalize=True, 
    reduce_lr=0.0, 
    resume='', 
    resume_options=False, 
    rnn_layers=1, 
    rnn_type='lstm', 
    save_best_only=True, 
    save_for_predict=False, 
    seed=1023, 
    self_attention_opt=1, 
    task_name='QuAC', 
    train_dir='QuAC_data/', 
    tune_partial=400, 
    use_dialog_act=False, 
    use_elmo=True, 
    use_ner=True, 
    use_pos=True, 
    use_wemb=True, 
    weight_decay=0

result:

    Train:  F1 = 3.298,     NA = 100.000,   TH = -18.768
    Test:   F1 = 3.30,      HEQ_Q=2.74,     HEQ_D = 0.81

CoQA

parameters:

gpu1 coqa_perq0:
    CoVe_opt=1, 
    MTLSTM_path='glove/MT-LSTM.pth', 
    answer_type_num=4, 
    batch_size=1, 
    concat_rnn=False, 
    cuda=True, 
    deep_att_hidden_size_per_abstr=250, 
    deep_inter_att_do_similar=0, 
    dev_dir='CoQA/', 
    do_hierarchical_query=True, 
    do_not_save=False, 
    do_prealign=True, 
    do_ptr_update=True, 
    do_residual=False, 
    do_residual_everything=False, 
    do_residual_rnn=False, 
    do_seq_dropout=True, dropout_emb=0.4, 
    elmo_batch_size=12, 
    elmo_lambda=0.0,
    epoches=30, 
    eval_per_epoch=1, 
    explicit_dialog_ctx=1, 
    final_output_att_hidden=250, 
    fix_embeddings=False, 
    grad_clipping=10, 
    hidden_size=125, 
    learning_rate=0.1, 
    log_file='output.log', 
    log_per_updates=20, 
    max_len=15, 
    model_dir='models', 
    momentum=0, 
    my_dropout_p=0.4, 
    name='', 
    ner_dim=8, 
    ner_size=19, 
    no_dialog_flow=False, 
    no_em=False, 
    optimizer='adamax', 
    pos_dim=12, 
    pos_size=51, 
    prealign_hidden=300, 
    prealign_option=2, 
    pretrain='', 
    ptr_net_attn_type='Bilinear', 
    ptr_net_indep_attn=True, 
    question_merge='linear_self_attn', 
    question_normalize=True, 
    rationale_lambda=0.0, 
    reduce_lr=0.0, 
    resume='', 
    resume_options=False, 
    rnn_layers=1, 
    rnn_type='lstm', 
    save_best_only=True, 
    save_for_predict=False, 
    seed=1023, 
    self_attention_opt=1, 
    task_name='CoQA', 
    train_dir='CoQA/', 
    tune_partial=400, 
    use_elmo=True, 
    use_ner=True, 
    use_pos=True, 
    use_wemb=True, 
    weight_decay=0

result:

    Train:  F1 = 4.599,
    Test:   F1 = 4.605

Environment

Tensorflow1.12 + CUDA9.0 + python3.6 + Tensor2tensor

Download and install Anaconda

You can find more install details about this on my blog.

sudo bash Anaconda3-5.2.0-Linux-x86_64.sh

conda --version

conda create -n tensorflow pip python=3.6

source activate tensorflow

pip install --upgrade pip

(tensorflow)$ pip install --ignore-installed --upgrade https://download.tensorflow.google.cn/linux/gpu/tensorflow_gpu-1.8.0-cp36-cp36m-linux_x86_64.whl

Encoding

Bert

Download pre-trained encoder model: BERT_4_Chinese

If you want to extract your own proposal data, you could just run the utils/prepare.py (some paths of files need to be modified:

source activate tensorflow
pip install -r requirements.txt
python prepare.py

GloVe

You could find our modified glove for Chinese. glove-tools

If you want to extract your own proposal data, you could find more details from glove-tools

Due to lack of support for Chinese in basic projects, we adopted LTP, LTP_py

LTP can be download from Google_Drive

CoVE

CoVE project: cove

ELMo

Bidirectional LSTM-CRF and ELMo project: anago

Whole project files structure

.
├── baseline
│   ├── DrQA
│   │   ├── download.sh
│   │   ├── drqa
│   │   │   ├── __init__.py
│   │   │   ├── pipeline
│   │   │   │   ├── drqa.py
│   │   │   │   └── __init__.py
│   │   │   ├── reader
│   │   │   │   ├── config.py
│   │   │   │   ├── data.py
│   │   │   │   ├── __init__.py
│   │   │   │   ├── layers.py
│   │   │   │   ├── model.py
│   │   │   │   ├── predictor.py
│   │   │   │   ├── rnn_reader.py
│   │   │   │   ├── utils.py
│   │   │   │   └── vector.py
│   │   │   ├── retriever
│   │   │   │   ├── doc_db.py
│   │   │   │   ├── __init__.py
│   │   │   │   ├── tfidf_doc_ranker.py
│   │   │   │   └── utils.py
│   │   │   └── tokenizers
│   │   │       ├── corenlp_tokenizer.py
│   │   │       ├── __init__.py
│   │   │       ├── regexp_tokenizer.py
│   │   │       ├── simple_tokenizer.py
│   │   │       ├── spacy_tokenizer.py
│   │   │       └── tokenizer.py
│   │   ├── img
│   │   │   └── drqa.png
│   │   ├── install_corenlp.sh
│   │   ├── LICENSE
│   │   ├── PATENTS
│   │   ├── README.md
│   │   ├── requirements.txt
│   │   ├── scripts
│   │   │   ├── convert
│   │   │   │   ├── squad.py
│   │   │   │   └── webquestions.py
│   │   │   ├── distant
│   │   │   │   ├── check_data.py
│   │   │   │   ├── generate.py
│   │   │   │   └── README.md
│   │   │   ├── pipeline
│   │   │   │   ├── eval.py
│   │   │   │   ├── interactive.py
│   │   │   │   └── predict.py
│   │   │   ├── reader
│   │   │   │   ├── interactive.py
│   │   │   │   ├── predict.py
│   │   │   │   ├── preprocess.py
│   │   │   │   ├── README.md
│   │   │   │   └── train.py
│   │   │   └── retriever
│   │   │       ├── build_db.py
│   │   │       ├── build_tfidf.py
│   │   │       ├── eval.py
│   │   │       ├── interactive.py
│   │   │       ├── prep_wikipedia.py
│   │   │       └── README.md
│   │   └── setup.py
│   └── FlowQA
│       ├── CoQA_eval.py
│       ├── download.sh
│       ├── general_utils.py
│       ├── predict_CoQA.py
│       ├── predict_QuAC.py
│       ├── preprocess_CoQA.py
│       ├── preprocess_QuAC.py
│       ├── QA_model
│       │   ├── detail_model.py
│       │   ├── layers.py
│       │   ├── model_CoQA.py
│       │   ├── model_QuAC.py
│       │   └── utils.py
│       ├── README.md
│       ├── requirements.txt
│       ├── train_CoQA.py
│       └── train_QuAC.py
├── bert
│   ├── chinese_L-12_H-768_A-12
│   │   ├── bert_config.json
│   │   ├── bert_model.ckpt.data-00000-of-00001
│   │   ├── bert_model.ckpt.index
│   │   ├── bert_model.ckpt.meta
│   │   └── vocab.txt
│   ├── extract_features.py
│   ├── __init__.py
│   ├── modeling.py
│   ├── __pycache__
│   │   ├── extract_features.cpython-35.pyc
│   │   ├── extract_features.cpython-36.pyc
│   │   ├── __init__.cpython-35.pyc
│   │   ├── __init__.cpython-36.pyc
│   │   ├── modeling.cpython-35.pyc
│   │   ├── modeling.cpython-36.pyc
│   │   └── tokenization.cpython-36.pyc
│   ├── run_classifier.py
│   ├── tmp_in.txt
│   ├── tmp_out.txt
│   ├── tokenization.py
│   └── tokenization_test.py
├── data
│   ├── chzhu.pkl
│   ├── dldi.pkl
│   ├── jwhu.pkl
│   ├── lmzhang.pkl
│   ├── xfduan.pkl
│   ├── xichen.pkl
│   ├── zh_session_ano.json
│   ├── zqzhu.pkl
│   └── zyzhao.pkl
├── __init__.py
├── README.md
├── requirements.txt
├── tree.txt
└── utils
    ├── __init__.py
    ├── PERQAInstance.py
    ├── prepare.py
    ├── __pycache__
    │   ├── PERQAInstance.cpython-35.pyc
    │   └── PERQAInstance.cpython-36.pyc
    ├── tmp_in.txt
    └── tmp_out.txt